Operations

The cost of deciding

The expensive part of an AI project is not the software. It is who has to think about it.

Ask a contract manufacturer what an AI project would cost and you get a number with a dollar sign on it. Software, integration, maybe a consultant. That number is the easy part, and it is almost never the part that hurts.

The expensive part is who has to think about it.

An AI plan worth acting on requires someone who knows how your plant runs. Not the IT contractor, not the vendor's solution architect. Someone who knows that the scheduling spreadsheet is the real master, that the second shift works around the label printer, and which of your customers will not accept an electronic record. At a company with 125 people, there are maybe three people like that, and all three are already fully committed to making this month's numbers.

So the cost is not the software. It is a claim on the attention of the people who keep production moving. That cost comes in three parts, and they get muddled together, which is why it is so often underestimated.

Three costs, not one

The first is the one everybody calculates. Your operations lead spends some hours on the project. You can put a number on those hours in about five minutes, and it looks affordable. This is the number that gets into the business case.

The second is what those hours were going to be spent on. This one is larger and nobody writes it down. When your best operator is in a discovery workshop, they are not on the floor solving the thing that would have stopped a changeover from running long. The work does not disappear, it gets done later, or by somebody less capable. In a plant, the cost of the second-best person handling a problem is rarely zero, and sometimes it is a held lot.

The third is the savings you did not capture while you were deciding. This is the one that compounds. Every month spent working out what to do is a month the burdensome process keeps running: the four hours a week someone spends reconciling two systems by hand, the rework a better check would have caught at the line instead of at final release, the corrections that ripple downstream into shipping and invoicing. Those costs do not pause while you deliberate.

The version where you have already started, and it still costs you

There is a variant of this that looks like progress. A leader decides AI is worth understanding and puts two or three capable people on it. Nobody is taken off the floor, exactly. They are just expected to work it out alongside everything else.

So they each go and muddle with it individually. One is pasting spec sheets into a chatbot, another is trying to get a report written, a third has a subscription nobody approved. All of it on their own time, none of it written down, and no two of them solving the same problem the same way. It feels like adoption. It is three people paying the learning cost separately and none of them getting to keep it.

McKinsey's 2025 survey of manufacturing COOs has a number that describes this: a quarter said they struggle to build applications that are reusable and scalable. Reusable is the word doing the work. Individual muddling produces nothing reusable, so the attention spent buys a skill that lives in one person's head and leaves when they are busy, or when they leave.

Why an owner skips the second cost, and I include myself

I skip it too. When I decide something is worth doing, the calculation I run is whether the thing is worth doing. Not what it costs to have my best person stop doing their job long enough to find out. Owners and builders are wired to look at the prize.

The displacement never shows up on a page because it never shows up on an invoice. Nobody sends you a bill for the changeover that ran long while your operations lead was in a workshop. There is no line item, so there is no number, so it does not enter the decision. It is the most expensive thing in the project and it is invisible by construction.

Which is why naming it is worth more than measuring it. There is no published figure for what your ops lead's redirected attention costs, and there will not be, because it depends entirely on what that person would otherwise have been doing this month. But an owner who has the thought before committing makes a different decision than one who has it in month four.

The data says this is a people problem

This is not a hunch. It is what the research keeps finding, from several directions.

McKinsey surveyed 101 manufacturing COOs in mid 2025. Two thirds said their companies were still at the exploration or targeted-implementation stage with AI. Two percent said AI was fully embedded across operations. Asked what was blocking them, half named the need to shift their culture, nearly half named reskilling, and 46 percent named limits in their data or IT and OT systems. Not one of the leading barriers was the technology itself.

One caveat on that survey, and we would rather say it than have you find it. Every respondent was at a company with over a billion dollars in revenue. Those are organizations with dedicated transformation teams, full-time data people, and budget lines for exactly this. If they are at two percent fully embedded, then a contract manufacturer with 125 people and no spare headcount is not behind. It is normal. The lesson is not that you are late. It is that the constraint is attention and capability rather than software, and that having more money does not make the constraint go away.

The Panorama Consulting 2026 ERP Report, drawn from 170 organizations between January 2025 and January 2026, points the same way. The median project ran nine months. More than a quarter came in over budget and almost a quarter over schedule. The leading cause of the schedule overruns was not technical. It was organizational: governance, change resistance, and friction in redesigning processes. All of which are ways of saying the project needed more of people's attention than anyone planned for.

And the attention is getting scarcer. The Manufacturing Institute and Deloitte projected in 2024 that manufacturing would need 3.8 million new employees by 2033, with as many as 1.9 million of those roles potentially going unfilled. In the same period, 65 percent of manufacturers told the National Association of Manufacturers that attracting and retaining talent was their primary business challenge. The people you would assign to figure out AI are the same people you cannot replace.

Now the trap

Here is where it turns from an expense into a mistake.

Once a company has spent six months of its operations lead's attention on a project, that spend becomes an argument. Not a good argument, but a powerful one. Nobody wants to stand up in front of the leadership team and say the last six months produced a plan we should not build. So the plan gets built. It gets built because it was expensive to produce, which is not a reason, and everyone in the room knows it is not a reason.

That is the sunk cost trap, and in a modernization project it is worse than usual for a specific reason: the thing you spent was not money, it was credibility. You asked your plant manager and your quality lead to take time away from the floor. They gave it. If the output gets shelved, the next time you ask them to spend attention on a systems project, the answer is a slower yes. Some operations only get to spend that credibility once.

The cost of a bad AI plan is not the plan. It is that you cannot easily go back and do it properly.

The compounding part

Delay does not sit still. It moves in one direction and it picks up speed.

A month of deciding is a month of the manual reconciliation still being manual. That is direct labor hours, and those hours come with an error rate. Errors found late cost more than errors found early, which is the oldest arithmetic in quality. A correction caught at the line is a few minutes. The same defect caught at final release is a held lot, a deviation, an investigation, and a conversation with a customer. Caught after shipment it is a complaint, and in food or supplements it can become a recall.

For food manufacturers there is a date attached to this now. The FDA's Food Traceability Rule under FSMA 204 carries a compliance date of 20 July 2028, extended by thirty months from the original January 2026 date and codified by Congress. That extension reads like relief. It is runway. The plants that use it to get their records in order will be fine. The ones that treat it as permission to wait will be doing the same work in 2027 with less time and a worse negotiating position with their vendors.

So what do you do instead

Not "move fast." We are not going to tell you that, because moving fast on a modernization project is how you end up with an expensive plan nobody follows.

The honest answer is narrower: separate the deciding from the doing, and put a hard boundary around the deciding.

The deciding is a bounded piece of work. It has a beginning, an end, and a deliverable. It needs a few days of your people's time, concentrated rather than spread across six months of meetings, and it needs someone who has done it before, so the months you would spend learning the terrain are months you do not have to spend.

What comes out is a map of where you stand: the systems you depend on, including the spreadsheets and the paper, because paper is usually where the work is. A read on whether your data can support the things you might want to do. A short list of the two or three projects worth doing first, and an honest note about the twelve that are not.

Then you decide, with the sunk cost kept small enough that "not yet" is still a real option.

That last part is the whole point. A diagnostic that costs three weeks is a diagnostic you can act on or ignore. A discovery process that costs six months of your operations lead is a decision you have already made without noticing.